addit / harness

Local telemetry

addit-harness can keep a metadata-only usage log on your own machine, so you can see which agents, skills and /dev-flow steps you actually use and how they end. It is off by default, it is never sent anywhere by the plugin, and you can read or delete every file. The privacy policy is the binding description; this page explains how it works.

On this page

Turn it on or off

Open /config and switch on the addit-harness option Local telemetry (telemetry_local, declared in the plugin’s userConfig, default false). Switch it off the same way. While it is off, every telemetry hook exits immediately without starting Python and nothing is written.

flowchart LR
    H["Hook event<br/>(session, agent, skill, workflow, doc write)"] --> L{"telemetry_local on?"}
    L -- no --> X["exit, nothing written"]
    L -- yes --> V["telemetry.py builds one record"]
    V --> C{"passes the contract?"}
    C -- no --> D["dropped"]
    C -- yes --> F["append one line to<br/>sessions/YYYY/MM/DD/{session}.jsonl"]

What is recorded

Every record is checked against hooks/telemetry-contract.json (schema 2.2) before it is written. The contract has no free-text field: a value is an enum member, a name from the plugin’s own inventory, a count, a boolean, a timestamp, an id that Claude Code assigned, or a hash. A record that does not fit is dropped.

  • Which addit-harness agents, skills, commands and /dev-flow workflows ran, and how they ended. Agent names outside addit-harness are recorded as “other” (except pr-review-toolkit:code-reviewer, which /dev-flow uses).
  • Gate verdicts: triage, intake brief, design review, plan approval, code review, QA, and the frontend FC-1 line-cap hook’s findings.
  • Counts and sizes: characters of an agent’s final report; size, lines, mermaid blocks and evidence labels of Markdown documents written under a docs/ folder; and the document kind (plan, solution, report and so on).
  • The names and sizes of the standard instruction files (CLAUDE.md, AGENTS.md) and of the plugin’s own rules when they load, and the setup sync state.
  • Timestamps, and the session, agent and workflow-run ids Claude Code assigns.
  • Salted hashes (truncated SHA-256) of your project folder path, of document paths under docs/, and of /dev-flow work-item names. The salt is a random value stored only in the log folder. These hashes are pseudonymous, not anonymous: anyone holding both the log and the salt can test a guess against them.

What is never recorded

Prompts, responses, code, file contents, the names or paths of your files and folders, command lines, error text, branch or repository names, and account, email or organisation details. To count lines and diagrams, the hook reads Markdown files written under docs/ and keeps only the numbers.

Where it lives

One folder, the plugin’s data folder in Claude Code: ~/.claude/plugins/data/<plugin-id>/telemetry/.

  • sessions/YYYY/MM/DD/<session-id>.jsonl: one JSON line per event.
  • state/: the salt and small per-session bookkeeping files.
  • export/<timestamp>/: anything you export (below).
  • README.md: a plain-text note on what the folder is and how to delete it.

Day folders older than 90 days are deleted when a session starts, and the oldest days are deleted first if the folder grows past 100 MB. To delete everything now, switch the option off and remove the folder.

The same plugin data folder, outside telemetry/, also holds one file that is not telemetry and is created whether the option is on or off: the empty one-time welcome marker onboarding/welcome-v1 (see Getting started). It holds nothing and is never read by the exporter.

Export a summary

/addit-harness:telemetry-export [--days N]

You run this yourself; Claude cannot invoke it on its own. It reads the last N days (default 30), re-validates every line, and writes two files to export/<timestamp>/ in the log folder:

  • data.json: KPIs plus one row per session (hashed session id, day, duration, agents, outcomes, workflows, document writes, FC-1 hits). Lines that fail validation are skipped and counted in invalid_lines.
  • report.html: the same data as a page that opens offline in any browser. It loads no external scripts, fonts or images.

KPIs include the share of shipped agents and rules actually used, the agent completion rate, the share of designed work items that were implemented, the share re-designed at a higher tier, plan-approval passes before implementation, plans and solutions within the deep length ceiling, in-place revisions, solution docs that list candidates before your idea, and frontend runs without an FC-1 hit. A KPI the log cannot compute shows n/a with the reason (for example, token cost and latency are left to Claude Code’s own OpenTelemetry export). With no log yet, every KPI reads n/a.

Optional publish to claude.ai

After the local export, the command asks whether to publish the aggregated KPIs (no per-session rows, ids or hashes) as a private Artifact on claude.ai. Only an explicit “yes” in that run publishes; the answer is not remembered. The upload goes to Anthropic’s hosted service under your account through Claude Code’s own Artifact tool, not through a connection the plugin opens, and that page is governed by Anthropic’s terms. If the Artifact tool is not available to you (for example with an API-key, Bedrock or Vertex login), nothing is published and the local path of artifact.html is printed instead.

Cost

With the option on, each hook event costs roughly 41 ms at the median and 55 ms at the 90th percentile in local measurements; it varies with your machine. The telemetry hooks fail open (any error exits quietly) and all but the session-end hook have a 5-second timeout, so a telemetry problem does not block your session. With the option off, a small shell launcher exits before any Python starts.

Next

Edit this page on GitHub ↗